Leveraging Deep Learning on Anomaly Detection in National Identity Number (NIN) Verification for National Identity Management Commission (NIMC), Nigeria

National identity verification systems are critical public infrastructure, yet remain vulnerable to sophisticated fraud strategies including automated bot attacks, synthetic identity construction, geolocation spoofing, and API request abuse. In Nigeria, the National Identity Management Commission (NIMC) manages over 100 million National Identification Number (NIN) records through a high-volume, multi-channel API-driven verification architecture. Existing rule-based monitoring systems are inadequate for detecting complex behavioral anomalies at this scale. This study proposes, implements, and evaluates a Hybrid Autoencoder–Long Short-Term Memory (AE-LSTM) deep learning framework for unsupervised anomaly detection in NIN verification transactions. A synthetic dataset of 50,000 transaction records incorporating six documented fraud categories at a 10% anomaly rate was generated to simulate realistic NIMC verification behavior. A nineteen-feature input representation was derived through structured feature engineering. The hybrid model fuses deep autoencoder reconstruction error with temporal behavioral scoring over sliding windows. Experimental results demonstrate accuracy of 95.0%, ROC-AUC of 0.9810, F1-score of 0.762, and recall of 79.6%, significantly outperforming Isolation Forest and rule-based baselines. These results confirm the viability of deep learning-based behavioral anomaly detection for securing national digital identity infrastructure in developing economies.